feat: v0.21.0 — token budget tracking and context window early warning

[Siddhant-K-code]

Closes #27

What

Adds token_budget.py to track cumulative input token usage against the model's context window limit and warn before exhaustion.

Context limits

Built-in limits for Anthropic (claude-*), OpenAI (gpt-4o, o1, o3), Google (gemini-1.5/2.0), and Meta (llama-3.1) models. Prefix matching handles versioned model names like claude-sonnet-4-20251022.

analyse_token_budget()

Per-request accumulation table:

Session: a1b2c3d4e5f6  Model: claude-sonnet-4  Context limit: 200,000 tokens
────────────────────────────────────────────────────────────────────────
  Req   Offset   Input tok  Output tok   Cumulative   % limit
────────────────────────────────────────────────────────────────────────
    1    +0:00      12,400       1,200       12,400     6.2%
    2    +1:34      18,600       2,100       31,000    15.5%
    3    +3:12     145,000       3,400      176,000    88.0%  ← warning
────────────────────────────────────────────────────────────────────────
Current: 176,000 tokens  ( 88.0% of limit)
Est. remaining: ~0 request(s)

TokenBudgetWatcher

Stateful watcher integrated into watch.py. Fires once when cumulative input tokens cross the threshold. Configurable via max_context_pct in .agent-watch.json or --max-context-pct CLI flag.

CLI

agent-strace token-budget [session-id] [--warning-threshold 0.9]

Tests

tests/test_token_budget.py — 13 tests.